Voice price wars intensify as ultra-low-bit compression breaks the two-bit barrier for edge agents.
The latest wave of releases brings aggressive cost competition to multimodal models and breakthroughs in local execution. xAI rolled out Grok Voice Transcribe 2.0 with leading streaming accuracy at aggressive pricing, while PrismML showed sub-2-bit weight compression on a 27B model, paving the way for full-grade open models running on consumer hardware without quality loss.
xAI ships Grok Voice Transcribe 2.0 with 6.8% multilingual WER
SourceAt $0.10/hour batch and $0.20 streaming, low-latency audio intelligence gets dramatically cheaper for voice-driven apps and conversational agents.
PrismML Bonsai 2 compresses 27B model to 5.9GB at 1.71 bits per weight
SourceBreaking the historical 2-bit quality cliff means developers can deploy 27B-class reasoning and agent models fully on-device on low-memory edge hardware.
Alibaba launches Qwen3.8-Omni-Flash at $0.47 output tokens
SourceProvides native omnimodal support (text, audio, image, video) across a 1M context window while slashing inference overhead for multimodal agents.
Get tomorrow's digest the moment it ships.